JP2013257668A - Interest analysis method, interest analyzer and program of the same - Google Patents

Abstract

PROBLEM TO BE SOLVED: To enable information recommendation precision to be enhanced by automatically extracting an optimal context condition.SOLUTION: A global context/context ID setting part 117 collects context conditions related to content browsing. A divided context extracting and processing part 116 divides an interest model into tables for combination, on the basis of the combinations of the collected context conditions, and extracts a table to be updated from the tables for each combination, on the basis of a relationship between the context conditions included in the combinations, to calculate a weight for the table to be updated on the basis of the relationship. An interest model updating and processing part 130 updates a user interest score with respect to a concept of the table to be updated by using a feature score calculated from content browsing history and the weight.

Description

  The present invention relates to an interest analysis method, an interest analysis apparatus, and a program for analyzing an interest of a user in consideration of a context by using meta information indicating a user's content browsing history and a concept representative of browsing target content.

  There is a demand for a technique for recommending appropriate services and contents in accordance with user actions and situations. For this reason, a technique for estimating a user's hobbies and the like from history information has been proposed, such as estimating a user's interest from a book information browsing history within the site and recommending a book. In such a method, assuming that meta information for summarizing contents is given to each content, a method for estimating the user's interest and the like from the frequency of concepts and the like appearing in the user history is a content-based filtering method ( Content Based Filtering (CBF) is being researched especially as a memory-based method.

  Specifically, the content-based filtering technique presents products of the same brand (holding the same concept tag) when browsing products of a specific brand (information indicating the brand holds the concept tag). In this case, the memory-based method presents a product of a specific brand if the specific brand is frequently browsed from the history of browsing in the past. In such a technique, there is a method for analyzing a history as a selection from candidate items. Also, with such a method, it is possible to handle context changes as learning weights and learning model switching (see, for example, Non-Patent Document 1 or 2).

Kenta Oku, Shinsuke Nakajima, Jun Miyazaki, Toshiaki Uemura, “Proposal of Context-Aware SVM-based Context-Aware Information Recommendation Method”, Database Society of Japan, DBSJ Letters Vol.5, No.1, pp.1-4, 2006 June Alexandros Karatzoglou, Xavier Amatriain, Linas Baltrunas, Nuria Oliver, Multiverse Recommendation: N-dimensional Tensor Factorization for Context-aware Collaborative Filtering, RecSys 2010: 79-86

  However, in the prior art, it was necessary to manually determine an appropriate classification in advance for the definition of the context. For this reason, it has been difficult to set an appropriate context. In addition, even when the context is set, there is a problem that if the setting is not appropriate, the history applicable to the context cannot be sufficiently collected and an appropriate recommendation result cannot be obtained.

  The present invention has been made paying attention to the above circumstances, and the object of the present invention is to automatically extract an optimal context condition and to improve the accuracy of information recommendation, an interest analysis method, an interest analysis device, and To provide that program.

  To achieve the above object, according to a first aspect of the present invention, an interest model having a user interest score for each of a plurality of concepts is used to analyze a user's interest from a browsing history of content including the concepts. A method, an apparatus, and a program, wherein a first content list browsing a plurality of contents as a list and a second content list browsing contents bodies from the first content list are clustered, and each cluster In addition, the total number of contents in the first content list is the first total number, the number of contents in which the concept appears in the first content list is the first appearance number, and the contents in the second content list The second total number and the second content list the number of contents in which the concept appears in the second content list. When the number of appearances is used, the number of contents in which the concept appears in the second content list under the conditions of the first total number, the first number of appearances, and the second total number is the second number. A first probability that is equal to or greater than the number of occurrences and a second probability that is equal to or less than the second number of occurrences, and a cumulative distribution function of a standard normal distribution based on the first and second probabilities The feature score is calculated by the inverse function of the above, the context conditions regarding the browsing of the content are collected, the interest model is divided into the table for each combination based on the combination of the collected context conditions, and included in the combination An update target table is extracted from the table for each combination based on the relationship between the context conditions to be calculated, a weight is calculated for the update target table based on the relationship, It is intended to update the user interest score for the concepts in the update target table using A and the said weight.

  According to the first aspect, it is possible to automatically extract context conditions from the way of collecting history data, and automatically determine the context suitability (weight) at the time of learning for the automatically extracted context conditions. Thereby, the processing cost and the operation cost can be reduced by the context condition automatic discrimination function, and the information recommendation can be made highly accurate by analysis based on an appropriate context condition in consideration of various contexts.

According to a second aspect of the present invention, in the first aspect, the interest model is divided into a table for each combination based on the amount of browsing history that matches the context condition.
According to the second aspect, by correlating the context condition with the history amount, it is possible to avoid the division of the interest model based on the context condition having a small corresponding history. Accuracy can be realized.

According to a third aspect of the present invention, in the first or second aspect, context conditions at the time of content request are collected, and an evaluation score for the content is calculated using a table that matches the context conditions. .
According to the third aspect, by calculating a content evaluation score using a user interest model that matches a user situation at the time of content request, it is possible to recommend content that matches the user's interest with high accuracy. Become.

  That is, according to the present invention, it is possible to provide an interest analysis method, an interest analysis device, and a program thereof that automatically extract optimum context conditions and enable information recommendation accuracy to be enhanced.

The whole system figure using the interest analysis device concerning this embodiment. The block diagram which shows the function structure of each apparatus of FIG. The figure which shows the outline | summary of the interest analysis process using browsing history. The figure which shows the outline | summary of the interest analysis process in the case of setting context conditions. The figure which shows an example of content request data. The figure which shows an example of content browsing operation on a client terminal. The figure which shows the data structural example of a list browsing content list. The figure which shows the data structural example of detailed browsing content. The figure which shows the data structural example of a presentation content list. The figure which shows an example of a content database. The figure which shows an example of a user interest score database. The figure which shows an example of the historical amount database according to context. The figure which shows an example of a context / relevance definition database. The figure which shows the processing flow of a log | history information receiving part. The figure which shows the processing flow of the interest table selection process part of learning object. The figure which shows the processing flow of a global context / context ID setting part. The figure which shows the processing flow of an interest model update process part. The figure which shows the data structural example of an analysis parameter list. The schematic diagram for demonstrating operation | movement of a characteristic score calculation part. The figure which shows the detail of a characteristic score calculation process. The figure which shows the detail of an interest model update process. The figure which shows the processing flow of a context log | history additional recording process part. The figure which shows the processing flow of division | segmentation context extraction. The figure which shows the process outline | summary of a context division | segmentation method. The figure which shows the specific example of the weight calculation process based on a context condition. The figure which shows the processing flow of a content request | requirement receiving part. The figure which shows the processing flow of a utilization interest table selection process part. The figure which shows the processing flow of a content evaluation process part. The figure which shows an example of a content score list. The figure which shows the detail of a content evaluation process.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
FIG. 1 is an overall system diagram using the interest analysis apparatus according to the present embodiment. This system includes a client terminal 200, a content server 300, and an interest analysis device 100. The client terminal 200 and the content server 300, and the content server 300 and the interest analysis device 100 are connected to each other via a communication network. The user acquires desired content from the content server 300 through a browsing operation on the client terminal 200 and presents the acquired content on the screen of the client terminal 200 for browsing.

  The client terminal 200 collects a content browsing history by a user operation, browses a plurality of content as a list, a list browsing content list (first content list), and a detailed browsing content list that browses the content body from the content list The (second content list) is transmitted to the content server 300 together with the terminal context information at the time of browsing. Note that the terminal context information includes, for example, measurement time and measurement results of other terminal-owned sensors such as position information, acceleration, earth axis sensors, and thermometers. The content server 300 transfers the list browsing content list and the detailed browsing content list together with the terminal context information to the interest analysis device 100 via the communication network.

  The interest analysis apparatus 100 analyzes a user's interest from a browsing history of content including the concept as meta information, using an interest model having a user interest score for each of the plurality of concepts. Specifically, using the list browsing content list and the detailed browsing content list, a feature score and a user interest score for each concept appearing in the content are calculated, and the user's interest is estimated. Based on this user interest score, the interest analysis apparatus 100 generates a list of contents sorted according to the user's interest (sorted presented content list) from the “presentation content list” received from the content server 300. To the content server 300.

FIG. 2 is a block diagram showing a functional configuration of each device in FIG. Each unit in FIG. 2 can be realized by, for example, a CPU (Central Processing Unit) of each device and a control program executed on a memory.
The interest analysis device 100 includes a history information receiving unit 110, a learning target interest table selection processing unit 113, a context history additional processing unit 115, a divided context extraction processing unit 116, a global context / context ID setting unit 117, and a content request receiving unit 121. , Use interest table selection processing unit 124, interest model update processing unit 130, context / relevance definition database 131, context-specific history amount database 132, interest score database 140, presentation content list reception unit 150, content database 160, content evaluation processing Unit 170 and sorted content score list transmission unit 180.

FIG. 3 shows an overview of the interest analysis process using the browsing history of the interest analysis device 100.
The history information receiving unit 110 receives the list browsing content list and the detailed browsing content list from the client terminal 200 via the content server 300. The list browsing content list is, for example, a list of content in which the user browses only the content titles in a list. The detailed browsing content list is a list of content that the user has viewed the content (details) of the content body. For example, in FIG. 3, the list browsing content list includes contents 1 to 8, and the detailed browsing content list includes contents 1, 3, and 4. In FIG. 3, the content indicated by the hatched pattern indicates that the concept B appears in the content 1, 6, 7, and 8.

  The interest model update processing unit 130 uses the list browsing content list and the detailed browsing content list to calculate a feature score (Z value to be described later) of each concept using a statistical model of concept selection. Further, for the interest model (user interest table for each context set ID) divided for each context set ID, which will be described later, the user interest score is updated using the characteristic score.

  The content evaluation processing unit 170 calculates the user's evaluation score for the content by probability combining using the user interest score of each concept appearing in the evaluation content. In the example of FIG. 3, the evaluation score of the evaluation content 1 is obtained using the user interest scores of the concepts E, F, and D that appear in the content 1.

  FIG. 4 is a diagram showing an overview of the interest analysis process when setting a context condition. The interest model update processing unit 130 automatically extracts a context set ID indicating a set of context conditions (context IDs) to be analyzed and a learning weight for the context set ID. This example shows a simple example. Specifically, context set IDs (24 sets) are prepared for each hour (same as a set for simplification but actually have only a time zone context ID), and weighted learning based on the browsing time To do. For example, if a browsing history occurs at 10:35, the interest table (MODEL 10) corresponding to this time and the interest tables nearby (for example, MODEL 9, 11) are updated simultaneously. The interest table indicates a set of records having a common context set ID.

The content evaluation processing unit 170 evaluates content using an interest table that matches the content request time. For example, when recommending at 12:10 as shown in FIG. 4, content evaluation is performed using the interest table (MODEL 12).
However, if there are a lot of contextual cuts (time, place, temperature, day of the week, season, etc.) that can be analyzed, dividing the interest model by all combinations requires enormous amounts of computation and necessary computer resources. Become. Further, it is very difficult for the operator to set “weight” for all combinations. Therefore, a method for automatically extracting optimum context conditions from the way of collecting history data and automatically determining the context suitability (weight) at the time of learning for each context set ID consisting of a combination of context conditions will be described later.

(Client terminal)
2, the client terminal 200 includes a history collection unit 210, a history information transmission unit 220, a content presentation unit 230, a content request transmission unit 240, and a terminal information collection unit 250.
The content request transmission unit 240 makes a content presentation request to the content server 300 in accordance with a user instruction (input). Specifically, content request data as shown in FIG. 5 is transmitted to the content server 300. For example, the content request data includes a client terminal ID (or user ID), a request time, and terminal context information. The request time may be added in the content server 300. The client terminal ID (or user ID) is a number uniquely assigned to each terminal (or user), and is an ID that matches a user ID in a user interest table of the interest score database 140 described later.

FIG. 6 shows an example of a content browsing operation by the user on the client terminal 200.
Based on the sorted presentation content list received from the content server 300, the content presentation unit 230 displays a list from the top of the sort order within the range allowed by the display screen size of the client terminal 200.

  In the example of FIG. 6, ten contents (contents 1 to 10) are displayed in a list. The content in the lower order of the sort order can be displayed by the list by the user's flick, scroll bar operation or the like. The list of contents actually displayed on the client terminal 200 in this way is referred to as a list browsing content list. That is, not all the contents in the sorted presentation content list are displayed on the client terminal 200, and thus are not necessarily included in the list browsing content list. When the user selects a title of each content from this list by clicking or the like, the main body (details) of the content of the selected title (contents 3, 5, and 6 in FIG. 6) can be viewed. The content whose details are browsed is included in the detailed browsing content list.

  As described above, the history collection unit 210 collects user operation histories and creates a list browsing content list and a detailed browsing content list. The history information transmission unit 220 transmits the list browsing content list and the detailed browsing content list created by the history collection unit 210 to the content server 300.

  FIG. 7 shows a data configuration example of the list browsing content list in the case of FIG. The list browsing content list has a cluster ID, a content ID, and a browsing time. The cluster is an identifier ("1" in FIG. 7) uniquely assigned to the list browsing content list and the detailed browsing content list. When the user browses the list browsing content displayed at another time (time zone), another cluster ID is given. The conditions for newly assigning the cluster ID under conditions other than the time include when there is no operation for a certain period of time while the list browsing content list is displayed, or when the browsing user (user ID) is switched, the list browsing content list On the other hand, when a narrow search is performed from the viewpoint of the content genre or the like, the browsing mode may be switched in other browsing applications. The content ID is an identifier uniquely assigned to each content of the list browsing content, and is assumed to match a value held in a content database 160 described later.

  FIG. 8 shows an example of the data structure of the detailed browsing content list in the case of FIG. The detailed browsing content list has a cluster ID, a content ID, and a browsing time, like the list browsing content list. The cluster ID is the same value as the list browsing content list (“1” in FIG. 8). In the detailed browsing content list, the content ID and browsing time are the identifier of the content (contents 3, 5, and 6 in FIG. 8) that the user has selected from the browsing content and browsed the content and the time when the content was browsed.

(Content server)
In FIG. 2, the content server 300 includes a content transmission processing unit 310, a sorted presentation content list reception unit 320, a presentation content list transmission unit 330, a presentation content list input unit 340, a history information transfer unit 350, and a content request transfer unit. 360 is provided.

The history information transfer unit 350 transfers the list browsing content list and the detailed browsing content list received from the client terminal 200 to the interest analysis device 100 via the communication network.
The presentation content list input unit 340 receives a presentation content list that lists contents to be presented to the client terminal 200 used by the user by the service operator. The presented content list transmission unit 330 transmits the input presented content list to the interest analysis apparatus 100 via the communication network.

  FIG. 9 shows a data configuration example of the presented content list. The presented content list has a content ID, a concept ID / relevance list, a content body, and a content registration time. The content ID is a unique ID assigned by the content server 300 to each content. The concept ID / relationship degree list stores a concept ID of a concept that appears in the content and a set of relevance W that indicates the degree of relevance between the concept and the content. The concept ID / relevance degree list is set in advance for each content. As a specific example, content 1 (sports article) includes {“baseball” concept ID = 1, relevance = 0.5}, { “Soccer” concept ID = 2, relevance = 0.8}, {“golf” concept ID = 3, relevance = 0.6}...

  The concept ID matches the value stored in the interest score database 140. For example, the relevance is a value from 0 to 1, and the larger the relevance, the stronger the relevance. As the relevance, a value set by the service operator at the time of content registration or a value calculated by another system is used.

The sorted presentation content list receiving unit 320 receives a sorted presentation content list and a client terminal ID (or user ID) obtained by sorting a part or all of the presentation content list from the interest analysis device 100. The content transmission processing unit 310 transmits the sorted presentation content list to the client terminal 200 corresponding to the client terminal ID (or user ID).
The content request transfer unit 360 transfers content request data (FIG. 5) that is a content presentation request from the content request transmission unit 240 of the client terminal 200 to the interest analysis device 100.

(Interest analysis device)
Next, the detail of each part of the interest analysis apparatus 100 is demonstrated.
[Content database 160]
FIG. 10 shows an example of the data structure of the content database 160. The content database 160 has a content table.

The content table stores a content ID, a concept ID / relevance list, a content body, and a content registration time. The value received by the presented content list receiving unit 150 is stored.
[Interest score database 140]
FIG. 11 shows an example of the data structure of the interest score database 140. The interest score database 140 has a user interest table.

  The user interest table stores values of concept ID, context set ID, user ID (client terminal ID), TotalZ (user interest score), X, and Y. The definition and calculation method of TotalZ, X, and Y will be described later. That is, the data for each context set ID in the user interest table is created for each combination (set) of context conditions for each user (terminal). This is to perform an analysis assuming that the user takes a characteristic action when the context is combined.

[History database 132 by context]
FIG. 12 shows an example of the data structure of the history amount database 132 by context. The context-specific history amount database 132 includes a context-specific user history amount table and an analysis target context set table.

  The context-specific user history amount table includes a client terminal ID (user ID), a context ID, and a matching cluster ID list. The matching cluster ID list has the same value as the cluster ID of the list browsing content list in FIG. 7 and is given by the history information receiving unit 110 described later. The compatible cluster ID list indicates a group of browsing history cluster IDs that match the context ID (context condition).

  The analysis target context set table includes a user ID (client terminal ID), a context set ID, a matching context ID list, and an adjacent context set ID list. The matching context ID list indicates a group of context IDs of the context conditions that should be satisfied by this context set. The adjacent context set ID list indicates a group of context set IDs to be learned simultaneously and the weight w at that time when learning is performed from the history of the context set. For example, {context set ID1, 0.1}, {context set ID2, 0.5},.

[Division context / relationship definition database 131]
FIG. 13 shows an example of the data structure of the division context / relevance definition database 131. The division context / relevance definition database 131 includes a context ID table and a context relevance table.

  The context ID table includes a context ID and a context condition. The context condition is, for example, 24 hours divided into 8 and 0 to 3 o'clock, 3 o'clock to 6 o'clock, etc., and the air temperature is divided into 10 to 0 degrees or less, 0 to 5 degrees,. . . Conditions such as 30 degrees or more, Monday, Tuesday,. . . , Including conditions for Sundays, or seasonal and weather conditions such as spring, summer, autumn and winter, sunny, rainy, cloudy.

  The context relevance table includes a context ID, a related context ID, and a distance between the related context IDs. The context ID is a number uniquely assigned by a global context / context ID setting unit 117 described later. The related context ID is a group of context IDs having a relationship with this context ID. As for the distance to the related context ID, a numerical value indicating the distance between the context ID and each related context ID is stored as a set of the context ID and the distance value, and the larger the number, the more the relationship with the context ID. It shows that there is little nature.

[Presentation content list receiving unit 150]
The presented content list receiving unit 150 receives the presented content list as shown in FIG. 9 from the content server 300 and stores it in the content database 160 shown in FIG.
[History information receiving unit 110]
FIG. 14 shows a processing flow of the history information receiving unit 110.

(Receiving history data)
The history information receiving unit 110 receives a client terminal ID (or user ID), a list browsing content list, a detailed browsing content list, browsing time information, and each measurement time from the history information transfer unit 350 of the content server 300 via a communication network. The terminal context information (position information, acceleration, ground axis sensor, thermometer, etc., the detection result of other sensors owned by the terminal) that is paired with the information is received. Then, a cluster ID (unique value) is assigned and output to the learning target interest table selection process 113 (A-1).

[Learning Object Interest Table Selection Processing Unit 113]
FIG. 15 shows a processing flow of the learning target interest table selection processing unit 113. The learning target interest table selection processing unit 113 receives the cluster ID, client terminal ID (or user ID), list browsing content list, detailed browsing content list, browsing time information, and terminal context information from the history information receiving unit 110. (A-1).

(Select user interest table for learning)
First, in order to acquire the user's context (situation) at the time of history reception, the interest table selection processing unit 113 to be learned includes a cluster ID, a client terminal ID (or user ID), browsing time information, and terminal context information. To the global context / context ID setting unit 117 (A-2), the client terminal ID (or user ID) from the global context / context ID setting unit 117 and the context that matches the user situation and the global context at the time of history collection The ID group is received (A-3). At this time, there is a method of using the current adaptation of the server.

  The learning target interest table selection processing unit 113 obtains a list of context set IDs and w values that match the context conditions from the analysis target context set table of the context-specific history amount database 132 when the context IDs are obtained. In addition, the “other” context condition is always an analysis target, and this weight is set to an initial value w = 1. In addition to the method of reading from the pre-calculated database, there is a method of calculating the w value by the “relationship weight calculation” process of the division context extraction processing unit 116 starting from the input context condition.

  The learning target interest table selection processing unit 113 includes a client terminal ID (or user ID), a list browsing content list, a detailed browsing content list, a context target ID of the learning target user interest table, and a context set ID of the learning target user interest table. Each w value is output (A-4).

[Global Context / Context ID Setting Unit 117]
The global context / context ID setting unit 117 uses the input (A-2) from the interest table selection processing unit 113 to be learned or the input (C-2) from the use interest table selection processing unit 124 as a process activation trigger.

(Global context collection)
The global context / context ID setting unit 117 receives the client terminal ID (or user ID), browsing time information (time information in the case of C-2), terminal context information, and cluster ID (omitted in the case of C-2). When received, collect global context. For example, when the terminal context information includes position information such as GPS history, information such as temperature, humidity, and weather corresponding to the position information and time information is collected from the Internet. Or, collect event information that has become a hot topic in the world from the Internet, or collect information that is set by the operator in a timely manner. When the terminal context information includes position information such as GPS history, event information in the vicinity of the user position is collected from the Internet or the information set by the operator in a timely manner. Alternatively, update information of each user's Twitter, blog, etc., information such as season / day of week / holiday, etc., and other user profile information such as the user's age, sex, occupation, etc. according to presetting, etc. are collected.

(DB read)
Based on the collected context information group, the global context / context ID setting unit 117 reads a context ID group that matches the history collection time point (or server current time point) from the context / relevance definition database 131. In the case of A-2, the cluster ID and the context ID group suitable for the history collection time (or server current time) are output to the interest table selection processing unit 113 to be learned (A-3). In the case of C-2, a context ID group suitable for the history collection time (or server current time) is output to the utilization interest table selection processing unit 124.

[User interest model update unit 130]
FIG. 17 shows a processing flow of the user interest model update unit 130. The user interest model update unit 130 receives the client terminal ID (or user ID), the list browsing content list, the detailed browsing content list, the context set ID of the learning target user interest table, and the learning from the learning target interest table selection processing unit 113. The w value for each context set in the target user interest table is input (A-4).

(Appearance concept extraction)
The user interest model updating unit 130 extracts a concept ID that appears in each content in the list browsing content list and the detailed browsing content list of the cluster ID from the content database 160. Specifically, in FIG. 7 and FIG. 8, “concept ID” linked to each content ID is searched from the content table of the content database 160 of FIG. 10. The user interest model update unit 130 includes cluster data {cluster ID, list browsing content list, detailed browsing content list} and content ID / concept ID association list {{content ID, {related concept ID, ...}}, ...} and the appearance concept list {concept ID} are generated. The “content ID / concept ID association list” is a list of concept IDs searched based on the content ID. The “appearance concept list” is a list of all concept IDs of concepts that appear in each content included in the list browsing content list and the detailed browsing content list.

(Analysis parameter extraction)
The user interest model update unit 130 calculates the number of appearances for each concept in the “appearance concept list”, extracts analysis parameters necessary for calculating the feature score, and generates an analysis parameter list.
FIG. 18 shows a data configuration example of the analysis parameter list. The analysis parameter list includes, for each cluster ID, the total content S (first total) of the list browsing content list, the total content a (second total) of the detailed browsing content list, and the appearance concept list associated with the cluster ID. N and n are calculated for each concept ID. N (first appearance number) is the number of contents to which the concept ID is assigned in the list browsing content list. n (second appearance number) is the number of contents to which the concept ID is assigned in the detailed browsing content list. It should be noted that N and n are calculated for all concept IDs in the appearance concept list including the added superordinate concept.

  FIG. 19A shows a schematic diagram of the analysis parameter extraction process. For example, the case where the user browses the details of 10 (= a) contents from a list of 50 (= S) contents is shown. Here, 15 articles (= N) containing the concept of “baseball” among the 50 contents displayed in a list are displayed, and the concept of “baseball” is included in 10 contents viewed by the user. This indicates that there are five (= n) contents including “”.

(Feature score calculation)
The user interest model update unit 130 calculates a feature score Z for each concept ID using the analysis parameters S, a, N, and n. FIG. 20 shows details of the feature score calculation process. In FIG. 20, i is a conceptual identifier, and j is a cluster ID. When H1 (first probability) is the total number S of the list browsing contents included in the list browsing content list and the number N of the contents of the list browsing content where the concept i appears, a detailed browsing content is randomly selected and viewed. In this case, the cumulative probability that the number of detailed browsing contents in which the concept i appears is n or more. When H2 (second probability) is the total number S of the list browsing contents included in the list browsing content list and the number of contents N in which the concept i appears in the list browsing contents, a detailed browsing content is randomly selected and viewed. In this case, the cumulative probability that the number of detailed browsing contents in which the concept i appears is n or less. In the present embodiment, the cumulative probabilities H1 and H2 are obtained by the hypergeometric distribution, but are not limited to this method. Examples of other distributions include a binomial distribution and a normal distribution.

  As shown in FIG. 19B, for example, there are 5 or more contents including the concept of “baseball” among the 10 contents viewed by the user using the analysis parameters S, N, a, and n described above. It is shown that the probability of being “0.12”. Here, “0.12” corresponds to the value of the cumulative probability H1.

As an example of using the value of H2, consider the case where n is 0 in the above analysis parameters. In this case, the probability when the number of appearances is 0 or less is calculated. Specifically, in FIG. 19B, since the horizontal axis is the value of the item of 0, “0.02”.
Then, as shown in FIG. 20, the user interest model update unit 130 calculates the feature score Z by using the inverse function of the standard normal distribution cumulative distribution function using the calculated cumulative probabilities H1 and H2. As shown in FIG. 19C, the feature score Z is obtained by the inverse function of the cumulative distribution function of the standard normal distribution with H1 as the cumulative probability. When H2 is used as the cumulative probability, the feature score Z is obtained with the sign of the return value of the inverse function of the standard normal distribution cumulative distribution function being negative. The user interest model update unit 130 outputs an update target concept list {cluster ID, {concept ID, feature score = Z, weight = w},. The weight w is a w value for each update target interest table ID.

(DB update)
The interest model update processing unit 130 updates the user interest score (TotalZ) of each concept ID in the “update target concept list”. FIG. 21 shows details of processing of the interest model update processing unit 130. The interest model update processing unit 130 uses the user interest score update formula for each concept i shown in FIG. 21 for the concept (appearance concept) that appears in the content, and the user interest score TotalZ _in and X _{i (n−1).} , Y _{i (n−1)} is obtained. And about each record (TotalZ, X, Y) stored in the column corresponding to the concept ID and the client terminal ID (user ID) for the record corresponding to the context set ID of the user interest table of the interest score database 140 of FIG. ).

Here, X _{i (n−1)} is the sum of the squares of the weights w of the update target concept list in the past (up to the previous time) for each concept ID (represented by the identifier i here). Similarly, Y _{i (n−1)} is the sum of multiplication of the weight w of the past update target concept list and the feature score Z for each concept ID (represented by identifier i here).

  These X and Y hold intermediate results in the user interest score (TotalZ) calculation process. When giving priority to memory saving / storage, at least three variables of TotalZ, X, and Y are used as variables of each concept. This can be realized by holding a numerical value. When priority is not given to memory saving / storage, all the calculated concept scores and feature scores Z of the respective clusters are stored. In this case, storage of X and Y is not necessary.

In FIG. 21, n is an identifier indicating how many times the update process is performed. A series of processes for obtaining the user interest score TotalZ is performed in units of cluster IDs. When the unit in which this series of processes is performed is counted once, n is the number of times this series of processes has been performed. Is an identifier. i is an identifier of a concept ID. Z _in is a Z value used for each update process of concept i. Note that Z _ij is a Z value for each of the list browsing content list and the detailed browsing content list, and has a relationship of Z _ij εZ _in . Weight w _in is a weight to be used in each process of updating the concept i, is a value that is set in the updated concept list.

[Context history addition processing unit 115]
FIG. 22 shows a processing flow of the context history appending processing unit 115. The context history appending processing unit 115 has a client terminal ID (or user ID) from the interest table selection processing unit 113 to be learned, a context ID group suitable for the history collection time point (or the current time at the server), and a cluster ID. Input (B-1).

(DB update)
The context history addition processing unit 115 updates the context-specific user history amount table of the context-specific history amount database 132 based on the input information. Specifically, the input cluster ID is added to the column of the matching cluster ID list corresponding to the client terminal ID (or user ID) and context ID.
The context history record processing unit 115 outputs the client terminal ID (or user ID) and the context ID group that matches the history collection time point (or the current time at the server) to the divided context extraction processing unit 116 (B-2). .

[Division context extraction processing unit 116]
FIG. 23 shows a processing flow of the division context extraction processing unit 116. The divided context extraction processing unit 116 receives the client terminal ID (or user ID) from the context history addition processing unit 115 and a context ID group that matches the history collection time point (or the current time at the server) (B-2). ).

(History threshold judgment for each context condition combination)
For each client terminal ID (or user ID), the divided context extraction processing unit 116 refers to the context-specific user history amount table in the context-specific history amount database 132 and combines it with another context for each context ID to be updated. If the history amount exceeds a predetermined threshold, a combination in which the history amount exceeds the threshold is extracted. Here, since the user history amount table by context in FIG. 12 stores a matching cluster ID list for each context ID, the history amount for each context combination is calculated in common with the combined context ID. This is achieved by counting the number of associated cluster IDs.

  For example, in the case of a combination of contexts “8:00 to 10:00”, “near station”, and AND “weekdays”, a history amount (such as the number of browsing times) exceeding 10 is extracted as a threshold. When the extracted combination is not in the analysis target context set table of the context-based user history amount database 132, a context set ID is allocated as a new context set.

  FIG. 24 shows a processing outline of the context dividing method. Here, it is assumed that a two-dimensional context condition is set, and 12 times are prepared every two hours, and the position information ID is prepared as an area serial number obtained by dividing all regions of Japan by a grid. There are three existing context conditions: “4 o'clock to 6 o'clock” AND “location x”, “18 o'clock to 20 o'clock” AND “location y”, and “other” (initial context condition). , “6 o'clock to 8 o'clock” AND “place x” logs (history amount, number of browsing times, etc.) are additionally generated. When the log of “6 o'clock to 8 o'clock” AND “location x” becomes 5 or more which is a threshold value, the context condition of “6 o'clock to 8 o'clock” AND “location x” is divided. As a result, the context conditions are “4 o'clock to 6 o'clock” AND “location x”, “18:00 to 20:00” AND “location y”, “6 o'clock to 8 o'clock” AND “location x”, “ There are four "others" (initial context conditions). In FIG. 24, an example of a two-dimensional vector is shown to simplify the explanation, but in actuality, processing is performed using a multidimensional vector (time, place, temperature, weather, day of the week, season, etc.).

(Related context set ID extraction)
When the new context set ID and the context ID are input by the history amount threshold determination process, the divided context extraction processing unit 116 refers to the context relevance table of the context / relevance definition database 131 and creates a new context. A context ID in the set and a context ID having a relationship are extracted, and a context set ID having the extracted context ID is extracted. In the context ID extraction process having this relationship, the context relevance table of the context / relevance definition database 131 is regarded as a graph, and multiple hops are performed starting from the extraction target context ID. As the number of hops, “2 or 3 times” or “integrated value for each hop of the distance to the related context ID” is set as a threshold.

  “Total” and “arithmetic” of the context IDs in all context set IDs of the “integrated value for each hop of the distance to the related context ID” calculated above for the distance between the extracted context set ID and the new context set ID It is calculated by “average value”, “cosine similarity when considered as a vector, and reciprocal of Pearson distance”.

(Relationship weight calculation)
When the divided context extraction processing unit 116 receives the new context set ID, the context ID, the related context set ID group, and the distance between each context set ID by the related context set ID extraction process, A mutual relationship weight between the new context ID and the related context ID is calculated from the distance between the sets calculated in “related context set ID extraction”. At this time, the distance is normalized to a value of 0 to 1, using the distance as a weight. As a normalization method, there are a method such as a mean 0, a probability value in a standard normal distribution function whose variance is an operator set value, and a result in a case where a distance is a Z value in a cumulative probability value. In addition, there are a method using a logistic function and a method of dividing by a sufficiently large number.

(DB update)
The divided context extraction processing unit 116 updates the analysis target context set table of the context-specific history amount database 132 based on the new context set ID, the context ID, and the calculated relationship weight.
FIG. 25 shows a specific example of the weight calculation process based on the context condition. FIG. 25 shows an image for calculating the fitness (weight) of the context condition “6 o'clock to 8 o'clock” AND [place x] during learning. When the data of “2 o'clock to 4 o'clock” AND [location x] is generated as CASE 1, the distance from the context condition of “6 o'clock to 8 o'clock” AND [location x] is 2, so the weight function w = The learning process is executed on the history cluster added with the result of f (2) as a weight. When the data of “10 o'clock to 12 o'clock” AND [location 4] is generated as CASE 2, the distance from the context condition of “6 o'clock to 8 o'clock” AND [location x] is 3, so the weight function w = The learning process is executed on the history cluster added with the result of f (3) as a weight. Furthermore, when data of “6 o'clock to 8 o'clock” AND [location y] is generated as CASE 3, the distance from the context condition of “6 o'clock to 8 o'clock” AND [location x] exceeds the threshold value. (In this example, the threshold is 3), and the learning process is not executed.

  In other words, by automatically extracting the context, it is possible to optimize the context definition of what kind of situation (combination of situations) to analyze and to save labor of the definition work, and by correlating the context definition with the history amount Since it is possible to avoid context decomposition with few corresponding histories, it is possible to achieve high accuracy of information recommendation with minimum computer resources.

[Content Request Receiving Unit 121]
FIG. 26 shows a processing flow of the content evaluation processing unit 170. The content request receiving unit 121 receives content request data as shown in FIG. 5 including the client terminal ID (or user ID) from the content request transfer unit 360.

(Receiving content request history data)
The content request receiving unit 121 receives a client terminal ID (or user ID), time information, and terminal context information (position information, acceleration, ground axis sensor, thermometer, etc.) from the content request transfer unit 360 of the content server 300 via a communication network. , And the measurement time and measurement results of other sensors owned by the terminal. Then, the client terminal ID (or user ID), time information, and terminal context information are output to the utilization interest table selection processing unit 124 (C-1). There is also a method in which these pieces of information are further transferred to the history information receiving unit 110 and used for learning. There is also a method of collecting the history of the content request on the client terminal side, notifying the history request collecting unit 210 of the information from the content request transmitting unit 240, and transmitting the information to the interest analysis apparatus 100.

[Use interest table selection processing unit 124]
FIG. 27 shows a processing flow of the utilization interest table selection processing unit 124. The usage interest table selection processing unit 124 receives the client terminal ID (or user ID), time information, and terminal context information from the content request receiving unit 121 (C-1).

(Read conformance context ID)
The utilization interest table selection processing unit 124 outputs the client terminal ID (or user ID), time information, and terminal context information to the global context / context ID setting unit 117 (C-2), and the global context / context ID setting unit. From 117, a context ID group that matches the client terminal ID (or user ID) and the history collection time point (or the current time at the server) is received (C-3).

(Determine context set ID to be used (select user interest table))
Through the process of reading the matching context ID, a context ID group matching the client terminal ID (or user ID) and the history collection time (or the current time at the server) is input from the global context / context ID setting unit 117. Then, the utilization interest table selection processing unit 124 acquires a context set ID that satisfies the context condition from the analysis target context set table of the context-specific history amount DB 132 when the context IDs are collected. When a plurality of context set IDs are acquired, the context set having the largest number of context conditions is selected. Then, the use interest table selection processing unit 124 outputs the client terminal ID (or user ID) and the context set ID corresponding to the user interest table to be used (C-4).

[Content Evaluation Processing Unit 170]
FIG. 28 shows a processing flow of the content evaluation processing unit 170. The content evaluation processing unit 170 includes a presentation content list in a format as shown in FIG. 7 read from the content table of the content database 160, a client terminal ID (or user ID) from the usage interest table selection processing unit 124, and a user interest to be used. A context set ID for determining a record in the table is input.

(Score evaluation)
The content evaluation processing unit 170 evaluates the content from the presented content list using a record of the user interest table that matches the context set ID to be used. Then, the content evaluation processing unit 170 calculates an evaluation score for the content to be evaluated, and generates a content score list as shown in FIG. The content score list has a content ID, an evaluation score, a content body, and a content registration time.

FIG. 30 shows an example of an evaluation score calculation method. For example, the content evaluation formula shown in Figure 30, the user interest score TotalZ _i concept i evaluation score EntityZ _x for the evaluation content x, relevance _{W i} a content x and concepts i (or importance of concepts i), And a set p of concept IDs appearing in the content x. The concept identifier i corresponds to the concept ID in the set p.

The user interest score (TotalZ) used in the calculation of FIG. 30 is based on the client terminal ID (or user ID) from the user interest table (FIG. 11) of the interest score database 140 for the concept ID related to each content. Read out and use. In FIG. 30, when the content 1 in which the concept K, the concept B, and the concept D appear is an evaluation content, the evaluation score EntityZ _x = 0.18 using the TotalZ and the relevance W of the concept K, the concept B, and the concept D. And can be calculated. On the other hand, when the content 2 in which only the concept B appears is set as the evaluation content, the evaluation score EntityZ _x = −0.3 can be calculated using the TotalZ, W of the concept B. Content 1 with a large value of evaluation score EntityZ _x is displayed preferentially.

In addition, the evaluation score EntityZ _x can be obtained by the following methods 1 to 3.
As a first modification, it is obtained by EntityZ _x = MAX (TotalZ _i * W _i ). MAX (TotalZ _i * W _i ) is a function that returns the maximum value of TotalZ _i * W _i for i∈p.

The second modification, the value of EntityZ _x When the value of the MAX (TotalZ _i * _{W i)} exceeds the threshold value, the return value of _{MAX (TotalZ i * W i)} . MAX (TotalZ _i * W _i ) is a function that returns the maximum value of TotalZ _i * W _i for i∈p. When the threshold value is not exceeded, the result of the content evaluation formula in FIG. 30 is set to EntityZ _x . MAX () is a function that returns the maximum value of TotalZ _i * W _i with i∈p. The threshold is a value set by the service operator.

As a third modification example, only the value iεp where TotalZ _i is a positive value is extracted, and the value integrated by the content evaluation formula of FIG. 30 is EntityZ _x .
(Sort content in score order)
The content evaluation processing unit 170 sorts the content included in the content score list in descending order of the evaluation score EntityZ _x and outputs the sorted content score list to the content score list transmission unit 180.

[Sorted content score list transmission unit 180]
The sorted content score list transmission unit 180 transmits the sorted content score list and client terminal ID (or user ID) input from the content evaluation processing unit 170 to the content server 300 via the communication network.

  As described above, according to the above-described embodiment, the list of candidates for user selection is defined, and the number of appearances of concepts in content selection from there is analyzed, thereby taking into account the rarity of appearance of each concept. In addition, since it is possible to use history features of concepts that are not selected from the list, it is possible to estimate the user's interest with high accuracy.

In addition, it is possible to automatically extract context conditions (those that analyze trends by limiting the conditions) from the way of gathering historical data, and automatically determine the context suitability (weight) at the time of learning for the automatically extracted context conditions. It becomes possible.
When there are many context cuts (time, place, temperature, day of the week, season, etc.) that can be analyzed, the calculation amount for all combinations becomes enormous. Further, it is very difficult for the operator to set “weight” for all combinations. According to the method of the present embodiment, processing cost and operation cost can be reduced by the automatic context condition discrimination function, and information recommendation can be made highly accurate by analysis based on appropriate context conditions in consideration of various contexts.

Furthermore, by calculating a content evaluation score using a user interest table that matches a user situation at the time of content request, it is possible to recommend content that matches the user's interest with high accuracy.
Note that the present invention is not limited to the above-described embodiment as it is, and can be embodied by modifying the constituent elements without departing from the scope of the invention in the implementation stage. Further, various inventions can be formed by appropriately combining a plurality of constituent elements disclosed in the embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, you may combine suitably the component covering different embodiment.

  DESCRIPTION OF SYMBOLS 100 ... Interest analysis apparatus, 200 ... Client terminal, 300 ... Content server, 110 ... History information receiving part, 113 ... Interest table selection process part of learning object, 115 ... Context history addition process part, 116 ... Divided context extraction process part, 117 ... Global context / context ID setting unit 121 ... Content request receiving unit 124 ... Use interest table selection processing unit 130 ... Interest model update processing unit 131 ... Context / relevance definition database 132 ... history amount database by context 140 ... Interest score database, 150 ... Presented content list receiving unit, 160 ... Content database, 170 ... Content evaluation processing unit, 180 ... Sorted content score list transmitting unit, 210 ... History information transmitting unit, 220 ... History collecting unit 230 ... Content presentation unit, 240 ... Content request transmission unit, 250 ... Terminal information collection unit, 310 ... Content transmission processing unit, 320 ... Sorted presentation content list reception unit, 330 ... Presentation content list transmission unit, 340 ... Presentation content list Input unit, 350... History information transfer unit, 360... Content request transfer unit.

Claims (7)

A method of analyzing a user's interest from a browsing history of content including the concept using an interest model having a user interest score for each of a plurality of concepts,
Clustering a first content list browsing a plurality of contents as a list and a second content list browsing a content body from the first content list;
For each cluster, the total number of contents in the first content list is the first total number, the number of contents in which the concept appears in the first content list is the first appearance number, and the second content When the total number of contents in the list is the second total number and the number of contents in which the concept appears in the second content list is the second appearance number, the first total number, the first appearance number, And under the condition of the second total number, a first probability that the number of contents in which the concept appears in the second content list is greater than or equal to the second occurrence number and less than or equal to the second occurrence number. Calculating a second probability, and calculating a feature score by an inverse function of a cumulative distribution function of a standard normal distribution based on the first probability and the second probability;
Collecting contextual conditions for viewing the content;
Dividing the model of interest into a table for each combination based on the collected context condition combinations;
Extracting an update target table from the table for each combination based on the relevance between context conditions included in the combination;
Calculating a weight for the update target table based on the relevance;
And updating the user interest score for the concept in the update target table using the feature score and the weight.   The interest analysis method according to claim 1, wherein the interest model is divided into a table for each combination based on an amount of browsing history that matches the context condition.   The interest analysis method according to claim 1, further comprising a step of collecting context conditions at the time of requesting content and calculating an evaluation score for the content using a table that matches the context condition. Using an interest model having a user interest score for each of a plurality of concepts, an apparatus for analyzing user interest from a browsing history of content including the concept,
Means for clustering a first content list in which a plurality of contents are browsed as a list and a second content list in which content bodies are browsed from the first content list;
For each cluster, the total number of contents in the first content list is the first total number, the number of contents in which the concept appears in the first content list is the first appearance number, and the second content When the total number of contents in the list is the second total number and the number of contents in which the concept appears in the second content list is the second appearance number, the first total number, the first appearance number, And under the condition of the second total number, a first probability that the number of contents in which the concept appears in the second content list is greater than or equal to the second occurrence number and less than or equal to the second occurrence number. Calculating means for calculating a feature score by an inverse function of a cumulative distribution function of a standard normal distribution based on the first probability and the second probability;
Means for collecting contextual conditions relating to browsing of the content;
Means for dividing the model of interest into a table for each combination based on the combination of the collected context conditions;
Means for extracting an update target table from the table for each combination based on the relationship between the context conditions included in the combination;
Means for calculating a weight for the update target table based on the relevance;
An interest analysis apparatus comprising: means for updating the user interest score for the concept in the update target table using the feature score and the weight.   The interest analysis apparatus according to claim 4, wherein the interest model is divided into a table for each combination based on an amount of browsing history that matches the context condition.   6. The interest analysis apparatus according to claim 4, further comprising means for collecting context conditions at the time of requesting content and calculating an evaluation score for the content using a table that matches the context condition.   An interest analysis device program that causes a computer to function as each means constituting the interest analysis device according to any one of claims 4 to 6.

Images